Ali Moharrer

Extractable Common Randomness From Gaussian Trees: Topological and Algebraic Perspectives

In this paper, we study both topological and algebraic properties of unrooted Gaussian trees in order to characterize their security performance. Such performance is measured by the corresponding potential in extracting common randomness from a given tree, which is further determined by max-min and min-max conditional mutual information (CMI) values, subject to the order of selecting variables from the tree by legitimate nodes Alice and Bob, and an eavesdropper Eve, respectively. A new operation is proposed to transform a Gaussian tree into another, and also to order different Gaussian trees. Through such operation we construct several equivalent classes of Gaussian trees. Each class includes multiple Gaussian trees that can be partially ordered based on the associated max-min or min-max CMI metric, and thus, we can find the most secure and the least secure trees in each partially ordered set (poset). The union of all posets generates all possible non-isomorphic trees of the given number of variables. Then, we assign a particular polynomial to each Gaussian tree, and show that such polynomial can determine the relative security performance of the Gaussian tree with respect to other trees within the same class. In the end, based on a generalized integer partition method, we propose a novel approach to efficiently enumerate the most secure structures of all posets.

Evaluation of security robustness against information leakage in Gaussian polytree graphical models

Extensive works have been undertaken to develop efficient statistical inference algorithms based on graphical models. However, there still lacks sufficient understanding about how topological properties affect certain information related metrics for certain graphs. In this paper, we are particularly interested in finding out how topological properties of rooted polytrees for Gaussian random variables determine its security robustness, which is measured by our proposed max-min information (MaMI) metric. MaMI is defined as the maximin value of the conditional mutual information between any two random variables (nodes) in a given DAG, conditioned on the value of a third random variable, which is at full disposal of an eavesdropper, under a constraint of a given fixed joint entropy. We show some general topological properties which the desired max-min solutions satisfy. Under such properties, we prove the superior max-min feature of the linear topology for a simple but non-trivial case. The results not only help us understand the security strength of different rooted polytree type DAGs, which is critical when we evaluate the information leakage issues for various jointly Gaussian distributed measurements in networks, but also provide us another algebraic and analysis perspective in grasping some fundamental properties of such DAGs.

Synthesis of Gaussian trees with correlation sign ambiguity: An information theoretic approach

A new layered encoding scheme is proposed to effectively generate a random vector with prescribed joint density that induces a latent Gaussian tree structure. The encoding algorithm relies on the learned structure of tree to use minimal number of common random variables to synthesize the desired density, which we argue such algorithm is also computationally efficient. We characterize the achievable rate region for the rate tuples of multi-layer latent Gaussian tree, through which the number of bits needed to simulate such Gaussian joint density are determined. The random sources used in our algorithm are the latent variables at the top layer of tree along with Bernoulli sign inputs, which capture the correlation signs between the variables. In latent Gaussian trees the pairwise correlation signs between the variables are intrinsically unrecoverable. Such information is vital since it completely determines the direction in which two variables are associated. As a by-product of determining the achievable rate region, we quantify the amount of information loss due to unrecoverable sign information. It is shown that maximizing the achievable rate-region is equivalent to finding the worst case density for Bernoulli sign inputs where maximum amount of sign information is lost.

Topological and Algebraic Properties for Classifying Unrooted Gaussian Trees under Privacy Constraints

In this paper, our objective is to find out how topological and algebraic properties of unrooted Gaussian tree models determine their security robustness, which is measured by our proposed max-min information (MaMI) metric. Such metric quantifies the amount of common randomness extractable through public discussion between two legitimate nodes under an eavesdropper attack. We show some general topological properties that the desired max-min solutions shall satisfy. Under such properties, we develop conditions under which comparable trees are put together to form partially ordered sets (posets). Each poset contains the most favorable structure as the poset leader, and the least favorable structure. Then, we compute the Tutte-like polynomial for each tree in a poset in order to assign a polynomial to any tree in a poset. Moreover, we propose a novel method, based on restricted integer partitions, to effectively enumerate all poset leaders. The results not only help us understand the security strength of different Gaussian trees, which is critical when we evaluate the information leakage issues for various jointly Gaussian distributed measurements in networks, but also provide us both an algebraic and a topological perspective in grasping some fundamental properties of such models.

Successive synthesis of latent Gaussian trees

A new synthesis scheme is proposed to effectively generate a random vector with prescribed joint density that induces a (latent) Gaussian tree structure. The quality of synthesis is measured by total variation distance between the synthesized and desired statistics. The proposed layered and successive encoding scheme relies on the learned structure of tree to use minimal number of common random variables to synthesize the desired density. We characterize the achievable rate region for the rate tuples of multi-layer latent Gaussian tree, through which the number of bits needed to simulate such Gaussian joint density are determined. The random sources used in our algorithm are the latent variables at the top layer of tree, the additive independent Gaussian noises, and the Bernoulli sign inputs that capture the ambiguity of correlation signs between the variables.